Contextual Feedback Loops: Amplifying Deep Reasoning with Iterative Top-Down Feedback
This addresses the issue of reconciling high-level predictions with lower-level representations in deep learning models, offering a novel approach to improve performance across multiple domains.
The paper tackles the problem of conventional deep networks relying on one-way backpropagation by proposing Contextual Feedback Loops (CFLs), a lightweight mechanism that re-injects top-down context into earlier layers for iterative refinement, resulting in consistent gains on tasks like CIFAR-10, ImageNet-1k, SpeechCommands, and GLUE SST-2 with minimal overhead.
Conventional deep networks rely on one-way backpropagation that overlooks reconciling high-level predictions with lower-level representations. We propose \emph{Contextual Feedback Loops} (CFLs), a lightweight mechanism that re-injects top-down context into earlier layers for iterative refinement. Concretely, CFLs map the network's prediction to a compact \emph{context vector}, which is fused back into each layer via gating adapters. Unrolled over multiple feedback steps, CFLs unify feed-forward and feedback-driven inference, letting top-level outputs continually refine lower-level features. Despite minimal overhead, CFLs yield consistent gains on tasks including CIFAR-10, ImageNet-1k, SpeechCommands, and GLUE SST-2. Moreover, by a Banach Fixed Point argument under mild Lipschitz conditions, these updates converge stably. Overall, CFLs show that even modest top-down feedback can substantially improve deep models, aligning with cognitive theories of iterative perception.